Discord

Checklist AICC 2026 Round 10 · Task 1

What I Want and What I Don't

Predict one rectangular crop per image that contains the listed wanted objects and excludes the listed unwanted objects, learning from only 15 labelled examples.

  • Vision
  • Prompt-conditioned region selection (bounding-box prediction)
Read the official statement

The task

Each sample consists of an image and two lists of semantic labels: objects_a, which should lie inside the chosen rectangle, and objects_b, which should lie outside it. A label refers to the union of all instances of that class in the image. Every prompted label is present in the image, each prompted mask covers between 4% and 50% of the image area, and each list holds 3 to 4 unique labels with no label in both lists.

The training set is deliberately tiny: 15 labelled samples, each with the image, the prompts, the ground-truth box and binary semantic masks for all annotated labels (with a masks.json manifest). The test set has 150 samples with images and prompts only. Every training box contains more than 95% of each objects_a mask and less than 5% of each objects_b mask.

Contestants must predict one box [x1, y1, x2, y2) per test sample (x2 and y2 exclusive) and generalise from the 15 examples using only the two permitted pretrained models.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
dataset/train (15 samples: images/*.png, masks/<id>/*.png with masks.json, tasks.json with id, image, objects_a, objects_b, bbox) and dataset/test (150 samples: images/*.png and tasks.json without bbox).
You submit
CSV with columns id,bbox; bbox is a JSON array string of four integers [x1, y1, x2, y2] inside the image, one row per test id.
Scoring
Per sample, the coverage of each prompted label is the fraction of its mask inside the rectangle. TP = objects_a labels with coverage >= 0.8; FN = objects_a labels with coverage < 0.8; FP = objects_b labels with coverage > 0.2. F1 = 2TP / (2TP + FP + FN) (0.0 if the denominator is zero); the score is the mean F1 over test samples. Baseline 0.66, reference solution 0.89.
Rules
  • The only pretrained models allowed are DINO ViT-B/16 (facebook/dino-vitb16) and CLIP ViT-B/16 (openai/clip-vit-base-patch16).
  • Hand-labelling examples is prohibited.
  • Individual participation (maximum team size 1); at most 50 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 10, online on Kaggle, 28 Aug 2026 18:00 UTC – 30 Aug 2026 18:00 UTC.

Notebook

Competition

Solution

Details

Year
2026, Online (Kaggle)
Round
Round 10 · Task 1
Language
English
License
Varies by task: What I Want and What I Don't — MIT; Inverse Translation and Moment Match — CC BY-NC-ND 4.0 (Kaggle competition licences). Solutions repository: MIT., as stated by the source